Appropriateness of gdistsamp

243 views
Skip to first unread message

Lydia Beaudrot

unread,
Sep 4, 2013, 11:19:33 AM9/4/13
to unma...@googlegroups.com
Dear unmarked community,

My goal is to model variation in Bornean orangutan densities in response to changes in fruit availability over time. I would very much appreciate feedback on the appropriateness of using gdistsamp for this.

The research question is to what extent do orangutans track changes in food availability across the landscape. The data consist of distance sampling and phenology measurements at 12 sites along an elevational gradient collected over 64 months.

The environment is aseasonal, meaning there are no particular times of year in which fruit is more readily available. Instead, fruit availability is characterized by boom and bust cycles that occur over multiple years and are believed to be triggered by ENSO events. In addition, orangutans are very long-lived and have extremely slow life histories with adult females reproducing on average once every 6-8 years. Births and deaths during this 64-month period were therefore likely quite low.

My questions are: 

1) Would it be appropriate to set up the observation covariate matrix of M sites x T years as the 12 sites by 64 months?

2) If so, should I build temporal dependence between months as an availability covariate?

3) If sites are not fully spatially independent, how much might this affect the density estimates?

Thank you very much,
Lydia

Jeffrey Royle

unread,
Sep 4, 2013, 12:24:42 PM9/4/13
to unma...@googlegroups.com
hi Lydia,
 I think the gdistsamp models are appropriate for your situation.
 
 (1) yes
 
 (2) I think thats reasonable. Alternatively you could stack the replicate samples on top of each other and treat them as new "sites". If the animals are mixing over the landscape in response to food availability then the effective site (scale at which relevant covariates are measured) seems to me to be a "site x time" . The within site dynamics maybe aren't that important.
 
 (3) depends on how non-independent they are. I assume you mean that some individuals can show up in > 1 transect?  The effect of this non-independence is, I think, to produce some bias in standard errors which I'm guessing is really negligible if the overlap is not too great. The only way to address this explicitly is with simulation I think.
 
regards
andy
 
 


--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

Lydia Beaudrot

unread,
Sep 4, 2013, 2:17:12 PM9/4/13
to unma...@googlegroups.com
Hi Andy,

Thanks so much, this is very helpful. I think that your suggestion to stack the months might help clarify another question I have, which is the appropriate interpretation of the lambda and phi parameters from the following model:

mHC.1 <- gdistsamp(~ elevation, ~ No.stemsfruit, ~ replicates, data=OHunmarkedGDS, output="density", unitsOut="kmsq", mixture="NB", keyfun="hazard", K=500)

At this point I have been interpreting lambda as the density of orangutans “belonging to a site” and phi as the probability than an orangutan that belongs to a site is present in that site given the number of stems with fruit. Sine we believe that the animals move fairly freely about the landscape in response to food, would stacking the months allow us move us away from this interpretation and instead allow us to estimate densities without the constraint of the animals “belonging to a site”?

In terms of the spatial non-independence, yes, exactly the animals might show up in more than one site during a month because they were monitored on different days and the animals can reach all of sites within a few days.

Thanks again,
Lydia 

Jeffrey Royle

unread,
Sep 4, 2013, 7:17:58 PM9/4/13
to unma...@googlegroups.com
hi Lydia,
 I agree, the interpretation of phi is problematic because individuals are moving freely over the landscape. So the notion of "belonging to a site" is kind of nonsense.
 Stacking the data is therefore probably a better approach.
 Even with the dependence induced by individuals being at multiple sites, I think the inferences will be valid about "site x time" covariates such as local food availability.

I would recommend doing a simulation study where you evaluate the validity of inferences by a simulation study.

regards
andy



--

Lydia Beaudrot

unread,
Sep 4, 2013, 7:46:04 PM9/4/13
to unma...@googlegroups.com
Hi Andy, 

Thank you so very much. I really appreciate your help!

Lydia

Lydia Beaudrot

unread,
Sep 16, 2013, 2:21:50 PM9/16/13
to unma...@googlegroups.com

Hi again,

I have now analyzed the stacked orangutan data and have follow-up questions about error messages and estimates that seem too high.

An initial comparison of null models showed that the negative binomial distribution produced lower AIC values. I ran 10 models of interest using the NB distribution.

The default optim method produced the error message “Hessian is singular. Try using fewer covariates and supplying starting values” despite either using zero covariates. When I provided starting values, the error message “Error in optim(starts, nll, method = method, hessian = se, ...) : initial value in 'vmmin' is not finite” occurred. I was only able to get NB models (with 0-3 lambda covariates) to run by setting optim=SANN. The lambda estimates from the best-fit model were high (min: 2.6, median: 3.7, max: 49.2 orangutans/kmsq). Previous density estimates of orangutans at this site using the program DISTANCE range from a minimum of 0.5 to a maximum of 3.5 orangutans/kmsq.

Following the guidance of other forum posts, I then re-ran the 10 models using the Poisson distribution and the default optim method in the hopes that estimates would decrease, but instead they were extremely inflated (e.g. 200 orangutans/km sq).

My questions are:

1)    Might SANN be contributing to inflated estimates? If so, do you have any suggestions for dealing with the error terms caused by other optimization methods?

2)    What else might be causing estimates to be so high? It is possible that these animals are just too few and far between on the ground for the models to work well predicting their occurrence?

Thank you very much in advance,

Lydia

Reply all
Reply to author
Forward
0 new messages